Image color correction with machine learning

Abstract

Systems, methods and computer program products for image color correction are described herein. An embodiment includes identifying color candidates of mean color correspondences between a first image having desired color characteristics and a second image to which the desired color characteristics are to be applied, training a classifier to provide a metric that corresponds a degree of difference between the first image and the second image, and iteratively determining mean color correspondences between the first image and the second image using the metric as an objective function and generating a color-corrected image having the desired color characteristics of the first image using the determined mean color correspondences.

What is claimed is: 
     
       1. A computer-implemented method for image color correction, comprising:
 identifying color candidates of mean color correspondences between a first image having desired color characteristics and a second image to which the desired color characteristics are to be applied; 
 training a classifier to provide a metric that corresponds a degree of difference between the first image and the second image; and 
 iteratively determining mean color correspondences between the first image and the second image using the metric as an objective function, and 
 generating a color-corrected image having the desired color characteristics of the first image using the determined mean color correspondences, 
 wherein the identifying, the training, the iteratively determining and the generating are performed using one or more processors. 
 
     
     
       2. The method of  claim 1 , the training further comprising:
 recursively re-training the classifier to obtain updated metrics that correspond to a degree of difference between the first image and one or more color-corrected candidates of the second image; and 
 selecting a candidate image having the desired color characteristics after a pre-determined number of re-training recursions, the candidate image corresponding to a desired metric score. 
 
     
     
       3. The method of  claim 2 , further comprising:
 generating a transformation field in color space to transform one or more mean colors in the second image; and 
 generating a corrected image having the desired color characteristics of the first image using the transformed colors. 
 
     
     
       4. The method of  claim 2 , wherein the training further comprises:
 providing a low metric score for color-corrected candidates of the second image determined to have color characteristics different from the first image; and 
 providing a high metric score for color-corrected candidates of the second image determined to have color characteristics similar to the first image. 
 
     
     
       5. The method of  claim 3 , further comprising:
 measuring a color distance between two or more pairs of mean color correspondences between the first image and the selected candidate image; 
 identifying a pair from said two or more pails contributing to noise in the corrected image; and 
 removing the identified pair from said two or more pairs to reduce noise in the corrected image. 
 
     
     
       6. The method of  claim 3 , further comprising:
 applying one or more transformed colors to sparse pixel locations in the second image; and 
 propagating other remaining transformed colors in the second image to preserve color gradients occurring in the second image. 
 
     
     
       7. The method of  claim 1 , wherein the training comprises:
 resolving the first image and the second image into a virtual pyramid of overlapping image patches; and 
 down-sampling the image patches to determine features in each patch for the training of the classifier. 
 
     
     
       8. The method of  claim 1 , wherein the identifying comprises:
 identifying the color candidates using a second-order spectral correspondence method. 
 
     
     
       9. The method of  claim 1 , wherein the identifying further comprises:
 determining mean color values from the first image and the second image using a k-means algorithm. 
 
     
     
       10. The method of  claim 1 , wherein the identifying comprises:
 identifying a plurality of mean color values in the first image as color candidates to replace colors in the second image. 
 
     
     
       11. The method of  claim 1 , wherein the feature used for classifier training is a spatial pyramid matching kernel of LAB color space, where L represents color lightness and A and B represent color-opponent dimensions. 
     
     
       12. The method of  claim 1 , wherein the classifier is a support vector machine (SVM) classifier. 
     
     
       13. A processor-based system for image color correction, comprising:
 one or more processors; 
 a color candidate identifier configured to identify color candidates of mean color correspondences between a first image having desired color characteristics and a second image to which the desired color characteristics are to be applied; 
 a classifier configured to provide a metric that corresponds to a degree of difference between the first image and the second image; 
 an optimizer configured to iteratively determine mean color correspondences between the first image and the second image using the metric as an objective function; and 
 a color transfer module configured to generate a transformation field in color space to transform one or more mean colors in the second image and generate a color-corrected image having the desired color characteristics of the first image using the transformed colors, 
 wherein the color candidate identifier, the classifier, the optimizer and the color transfer module are implemented using the one or more processors. 
 
     
     
       14. The system of  claim 13 , wherein the classifier and the optimizer are configured to operate recursively. 
     
     
       15. The system of  claim 13 , wherein the metric indicates a higher score when the correctible image corresponds closely with the reference image, and a lower score when the correctible image does not correspond closely with the reference image. 
     
     
       16. The system of  claim 15 , wherein the optimizer is configured to measure the metric during an iterative correction process performed by the classifier and the optimizer on one or more one or more color-corrected candidates of the second image. 
     
     
       17. The system of  claim 16 , wherein the color transfer module corrects colors in the second image based on colors in the first image, wherein the metric comprises a maximum metric score. 
     
     
       18. A non-transitory Computer readable storage medium having instructions stored thereon that, when executed by a computing device, cause the computing device to perform operations comprising:
 identifying color candidates of mean color correspondences between a first image having desired color characteristics and a second image to which the desired color characteristics are to be applied; 
 training a classifier to provide a metric that corresponds a degree of difference between the first image and the second image; and 
 iteratively determining mean color correspondences between the first image and the second image using the metric as an objective function, and 
 generating a color-corrected image having the desired color characteristics of the first image using the determined mean color correspondences. 
 
     
     
       19. The computer-readable storage medium of  claim 18 , the training further comprising:
 recursively re-training the classifier to obtain updated metrics that correspond to a degree of difference between the first image and one or more color-corrected candidates of the second image; and 
 selecting a candidate image having the desired color characteristics after a pre-determined number of re-training recursions, the candidate image corresponding to a desired metric score. 
 
     
     
       20. The computer-readable storage medium of  claim 19 , the operations further comprising:
 generating a transformation field in color space to transform one or more mean colors in the second image; and 
 generating a corrected image having the desired color characteristics of the first image using the transformed colors. 
 
     
     
       21. A computer-implemented method for generating a satellite image map, comprising:
 obtaining a first image and a second image from one or more satellite imaging sources, wherein the first image has desired color characteristics and the second image in one to which the desired color characteristics are to be applied, and wherein the first image and the second image are partially overlapping images; 
 identifying color candidates of mean color correspondences between the first image and the second image; 
 training a classifier to provide a metric that corresponds a degree of difference between the first image and the second image; 
 iteratively determining mean color correspondences between the first image and the second image using the metric as an objective function; 
 generating a color-corrected image having the desired color characteristics of the first image using the determined mean color correspondences; and 
 merging the partially overlapping first image and the second image to generate a satellite image map, 
 wherein the obtaining, the identifying, the training, the iteratively determining, the generating and the merging are performed using one or more processors.